Sound-based sleep staging using pretrained speech foundation models

Sleep staging is essential for diagnosing sleep disorders and understanding sleep physiology, yet traditional polysomnography (PSG) is costly, intrusive, and impractical for large-scale or home-based monitoring. Wearable devices provide alternatives but face limitations such as motion artifacts and inconsistent skin contact. In this study, we propose a non-contact sleep staging approach using sound-based analysis. To address data scarcity, we employ transfer learning with pretrained speech foundation models, originally developed for automatic speech recognition but capable of capturing rich acoustic representations beyond linguistic content. By analysing temporal attention weights in the extracted HuBERT embeddings, we demonstrate that these models effectively capture respiratory patterns. Our findings suggest that repurposing speech foundation models for sleep staging provides a scalable, non-contact alternative to PSG, with promising applications in home-based and clinical settings.

Clinical Relevance—This demonstrates the feasibility of a non-contact, sound-based approach to sleep staging, potentially improving access and adherence in clinical and home settings.